MGML-ET: integrating TSEB-derived features with machine learning for large-scale evapotranspiration mapping

Evapotranspiration (ET) is a central process linking the terrestrial water, energy, and carbon cycles, underpinning climate regulation, ecosystem functioning, and water resource sustainability. However, reliable daily ET estimation remains challenging in regions with heterogeneous land cover, complex terrain, and limited observations conditions that are increasingly prevalent under climate change. Physically based models provide process interpretability but are sensitive to input data quality, while machine learning approaches offer strong predictive skill yet often lack physical consistency and robustness when extrapolated beyond observed conditions. Here, we develop a Mechanism-Guided Machine Learning framework for evapotranspiration estimation (MGML-ET), which introduces TSEB-derived process-informed information into a machine-learning framework as soft process guidance, rather than imposing hard physical constraints. The framework combines multi-source remote sensing, meteorological data, and TSEB-derived variables to produce daily ET estimates at 500 m resolution across China from 2001 to 2024. The optimal three-source configuration achieved strong predictive performance (testing R² = 0.85), while generally outperforming reduced-input configurations and providing greater robustness under stricter site-level validation. MGML-ET also reproduced realistic spatial heterogeneity and seasonal dynamics across major ecosystem types, although performance remained weaker in environmentally extreme settings such as deserts and wetlands. Long-term analysis further showed robust significant ET increases over 21.04% of China's land area, closely associated with vegetation recovery in major ecological restoration regions. These findings demonstrate that integrating process-informed information into data-driven models enhances the robustness and interpretability of ET estimation, while also highlighting the need for improved treatment of uncertainty in highly heterogeneous and climatically extreme environments.